Machine Learning for Predicting Early Functional Outcomes in Patients with Stroke

BACKGROUND: Development of a universal predictive tool for patients with stroke remains a challenge. AIM: The study aimed to develop machine learning–based models that could predict functional outcomes from the first day after stroke. The models were trained using clinical and anamnestic predictors, and functional outcomes were evaluated using the National Institutes of Health Stroke Scale (NIHSS) and the modified Rankin Scale (mRS) at hospital discharge. METHODS: Models based on artificial neural network (ANN) and random forest (RF) algorithms were developed using a database created from 5,225 records of patients with stroke discharged from neurological departments. Twenty-one parameters were used, including patient demographics; baseline National Institutes of NIHSS and mRS scores; stroke type; time from stroke onset to hospitalization; comorbidities; and emergency revascularization. Outcomes were predicted using NIHSS and mRS scores. The algorithms solved the classification problem for multiple sets of outcome values. Model I had 26 classes (NIHSS score pf 0–25), while model II had 6 classes (mRS score of 0–5). The quality of the models was evaluated using the area under the receiver operating characteristic curve (ROC-AUC). The contribution of each predictor was evaluated using the SHapley Additive exPlanations (SHAP). RESULTS: The predictive value of the ANN was determined based on the AUC-ROC: 0.771 for model I and 0.844 for model II. The RF AUC-ROC was 0.778 for model I and 0.845 for model II. The RF algorithm was chosen for further work due to its better interpretability. The most significant features that influenced the predicted outcomes were baseline NIHSS and mRs scores, patient age, time from stroke onset to admission, and stroke type. When RF performance was tested on an external validation set of 783 records, ROC-AUC values were 0.786 for model I and 0.774 for model II. A calculator was developed for practical use. CONCLUSION: The proposed RF-based models can reliably predict an early functional outcome within the first day of stroke onset, using NIHSS and mRs scores and clinical and anamnestic predictors. This tool can be used to develop personalized therapeutic and rehabilitation strategies in the acute phase of a stroke. These models are versatile to be used in rural and remote healthcare organizations that lack specialized staff and diagnostic equipment.